Research On Medical Knowledge Extraction In Electronic Medical Records Based On Deep Learning

With the application of Internet technology in the medical field,a large number of electronic medical records with rich medical knowledge have emerged.It implies a potential link between the patient's diagnosis and symptoms,medication status and treatment.The analysis and mining of these implicit knowledge can help clinical decision-making,and provide basic support for the subsequent establishment of a structured medical domain knowledge graph.However,the electronic medical record is an unstructured free text written by medical personnel,its expression is complex,including a large number of professional vocabularies in the medical field,and influenced by the writing habits of the writer.Therefore,the efficient and accurate of medical knowledge extraction contained in electronic medical records still faces enormous challenges.Based on natural language processing and deep learning technology,this paper deeply researched and discussed the tasks of named entity recognition and medical relation extraction in medical knowledge extraction,and proposes a highly usable implementation scheme to solve the existing problems of existing methods,such as relying on manual operations too much,sparse of features,poor applicability of methods,low efficiency of the models.This paper has made significant improvements in the extraction result.The main work of this paper is divided into the following three parts:(1)For the common distributed text representation method,this paper chose the GloVe model through experiment comparison,and used Wikipedia and I2B2 2010 English electronic medical record texts as the corpus of word vector training,and finally constructed the word vector of electronic medical record text.Moreover,this paper considered the problem that the word indicates the sparseness of the text information,and supplemented the letter information as a feature.The CNN-based letter feature extraction model was constructed.The output state of letter and word vector were spliced together as the input part of the model for named entity recognition and medical relationship extraction.(2)In the medical named entity identification task,this paper used the sequence labeling strategy as the solution.Considering that CNN can realize the advantages of local feature extraction,this paper applied it to the middle layer of word vector input and model classification layer,and proposed a named entity recognition model based on BiLSTM-CRF combined with inter-word feature convolution rules.The experiments showed that the proposed model achieved better results compared with the mainstream methods and verifies its effectiveness.(3)In the medical relation extraction task,aiming at the characteristics of the prior knowledge of the entity words,this paper proposed a medical relation extraction model based on the attention mechanism combined with BiLSTM-CRF,and verified its performance and extraction ability through experiments.Its result is far higher than other common methods.In addition,this paper considered the phenomenon that the error of the named entity recognition will be transmitted in the medical relation extraction.The entity candidate queue mechanism was proposed,and multiple candidates identified by the named entity were added to the medical relation extraction task in a queue form.The classification and extraction were completed according to the comprehensive performance of the two tasks.In summary,the task of named entity recognition and medical relation extraction in medical knowledge extraction has achieved better performance than previous research,and contributed to the basic research of medical knowledge extraction of electronic medical records.